The rise of generative AI has forced courts to examine copyright law in new ways. Two recent opinions – Judge Alsup’s ruling in Bartz v. Anthropic PBC and Judge Chhabria’s decision in Kadrey v. Meta Platforms – illustrate how judicial perspectives can lead to differing analyses of fair use, even when the outcomes appear aligned.
Both courts found that training AI on copyrighted books could qualify as fair use, but they diverged in critical ways, particularly regarding the nature of AI training, the use of pirated materials and the assessment of market harm. Fair use is an affirmative defence to copyright infringement, assessed under four factors codified in 17 U.S.C. § 107: (1) the purpose and character of the use, (2) the nature of the copyrighted work, (3) the amount and substantiality of the work used, and (4) the effect on the market for the work.
Both cases involved author-plaintiffs whose books were copied to train large language models (LLMs) – Anthropic’s ‘Claude’ in Bartz and Meta’s ‘Llama’ in Kadrey. Each defendant used unauthorised digital copies from ‘shadow libraries’. While both courts ultimately found fair use, their reasoning diverged. Both judges found that training on lawfully acquired materials is transformative (favouring factor one) and acknowledged that the highly creative nature of novels weighed against fair use under factor two. Even though entire books were used, both judges ruled that copying the full works was reasonable given LLM training requirements, thus favouring factor three.
The courts differed, however, in their treatment of pirated books leaving the impact of pirating copyrighted works for AI training uncertain. Alsup distinguished between pirated and legitimately acquired books. He ruled that using purchased books for training could qualify as fair use but held that Anthropic’s creation of a general-purpose library using pirated books could not.
Alsup viewed the initial acquisition of the training material as a separate act from using the materials in training and found that storing pirated books in a central library was “inherently, irredeemably infringing”. Chhabria, by contrast, was less concerned with how Meta obtained the books. He viewed all copies used by Meta as part of a transformative process, regardless of how they were acquired.
The most contested fair use factor – market harm – was where the opinions most diverged. Both judges considered three types of potential harm: (1) the development of a licensing market for training data, (2) reproduction of the original work, and (3) dilution from AI-generated substitutes. Both courts dismissed the licensing market argument, finding it to not be cognisable. They also rejected the claim that LLMs reproduced copyrighted works directly – Meta’s model, for example, could not reproduce more than 50 words of a book. However, the third harm – market dilution – exposed the biggest split.
Alsup was dismissive, likening the complaint to fears that teaching children to write might produce competition. Chhabria disagreed, calling the risk of LLM-generated substitutes unprecedented: “No other use... has anything near the potential to flood the market with competing works.”
He found for Meta on fair use primarily because the plaintiffs failed to argue or present sufficient evidence of market harm. But Chhabria warned that future plaintiffs could succeed by showing that AI-generated content harms markets for human-authored works.
He wrote that in cases involving uses like Meta’s “it seems like the plaintiffs will often win, at least where those cases have better-developed records on the market effects of the defendant’s use. No matter how transformative LLM training may be, it’s hard to imagine that it can be fair use to use copyrighted books to develop a tool to make billions or trillions of dollars while enabling the creation of a potentially endless stream of competing works that could significantly harm the market for those books”.
The US Copyright Office expressed concerns similar to Chhabria’s in its pre-publication report Copyright and Artificial Intelligence Part 3. Like Chhabria, the office rejected Meta’s argument that the fourth factor should focus only on direct market harm to specific works. The office also cautioned against schoolhouse analogies like the one used by Alsup, noting that AI learning differs from human learning in ways that are material to the copyright analysis.
Instead, the office emphasised that LLMs can generate content with volume and speed that threatens to crowd out human authors. Even if AI-generated work is of lesser quality, it can still dilute the market due to the limited time and attention of consumers. The office also noted that AI tools could mimic an author’s style so closely that audiences might be unable to distinguish between original and AI-generated works, effectively making authors compete with their own data.
These concerns extend to fields beyond literature. In music, AI-generated content could distort revenue by gaming algorithm-based royalty systems, as streaming platforms reward high-play volumes regardless of content origin.
The question of market dilution from non-infringing, yet competitive AI works will be a major question that courts will have to answer. The contrasting approaches in Bartz and Kadrey show that the legal boundaries of AI training remain fluid. While both courts reached similar conclusions, Chhabria left a clearer roadmap for future plaintiffs: build a strong record of market harm and you may prevail. The Copyright Office’s perspective suggests that even if market harm is not yet fully realised, courts should consider how widespread AI use might erode creative incentives.
On 14 July, Anthropic filed a motion for an interlocutory appeal, asking the judge to allow the United States Court of Appeals for the Ninth Circuit to review Bartz. If granted, an interlocutory appeal would mark only the second AI-related case to reach a federal appeals court, following Thomson Reuters v Ross Intelligence in the United States Court of Appeals for the Third Circuit. With courts already split on copyright and fair use, a Ninth Circuit ruling could rival the Third Circuit's in shaping the future of generative AI.
As generative AI continues to evolve, courts will soon face cases that force a definitive ruling on whether AI-induced market dilution defeats a fair use defence.
Avery Williams is a principal in McKool Smith’s intellectual property litigation practice. Joseph Micheli is an associate at McKool Smith specialising in patent litigation.
Email your news and story ideas to: [email protected]
